Gnat Robots - DSpace@MIT

MASSACHUSETTS INSTITUTE OF TECHNOLOGY
ARTIFICIAL INTELLIGENCE LABORATORY
June, 1987
Working Paper 295
Gnat Robots
(And How They Will Change Robotics)
A.M. Flynn
Abstract
A new concept in mobile robots is proposed, namely that of a
gnat-sized autonomous robot with on-board sensors, brains, actuators
and power supplies, all fabricated on a single piece of silicon. Recent
breakthroughs in computer architectures for intelligent robots, sensor
integration algorithms and micromachining techniques for building onchip micromotors, combined with the ever decreasing size of integrated
logic, sensors and power circuitry have led to the possibility of a new
generation of mobile robots which will vastly change the way we think
about robotics.
Forget about today's first generation robots: costly, bulky machines
with parts acquired from many different vendors. What will appear
will be cheap, mass produced, slimmed down, integrated robots that
need no maintenance, no spare parts and no special care. The cost
advantages of these robots will create new worlds of applications.
Gnat robots will offer a new approach in using automation technology. We will begin to think in terms of massive parallelism: using
millions of simple, cheap, gnat robots in place of one large complicated
robot. Furthermore, disposable robots will even become realistic.
This paper outlines how to build gnat robots. It discusses the
technology thrusts that will be required for developing such machines
and sets forth some strategies for design. A close look is taken at the
tradeoffs involved in choosing components of the system: locomiotion
options, power sources, types of sensors and architectures for intelligence.
A.I. Laboratory Working Papers are produced for internal circulation, and may contain information that is, for example, too preliminary or too detailed for formal publication. It is not intended that they should be considered papers to which reference can be
made in the literature.
1
Where Did All The Robots Go?
If you've been keeping up with your reading of Time magazine or watching of the Saturday morning cartoons, you were probably disappointed last
Christmas when you didn't get a robot that did the dishes, washed the windows and swept the floors. Where did all the robots go? We've become
conditioned to believe that soon robot-helpers would be permeating our society, but it hasn't turned out that way. Why don't we see more robots in
everyday life?
The main reason is money. Robot technology is very expensive for the
level of intelligence attainable. Many hard problems need to be solved in
sensory perception and intelligent control before robots will achieve higher
levels of competence. No market exists today for such costly machines of
limited capabilities. Therefore I propose that we work on building very cheap
robots with the capabilities we can produce now and then see what happens
later, much the same way as when microprocessors were first introduced (as
video games, etc.).
What makes robots expensive? Mobile robots today contain mostly motors and batteries while all the sensors and computers come in a very tiny
package. The battery-motor system has a certain runaway characteristic.
Big motors tend to need big batteries which weigh down the chassis, so
larger motors are called for, which require heftier batteries ... and on and
on it goes. Meanwhile, all the intelligence and sensing mechanisms fit onto a
few square inches of silicon. Mobile robots that are used as sensor platforms,
exploration robots or sentries, as opposed to heavy lift arm-type robots, pay
especially heavy penalties for carrying around large loads of motors and
batteries.
If mobile robots are half motors and batteries, what takes up the other
half of the physical space? The answer is connectors: power connectors,
signal connectors, bus interfaces - whatever it takes to hook up one vendor's computer to another vendor's motor to another vendor's sensor to yet
another vendor's battery. All these interfaces between parts from various
suppliers mean added cost and complexity and the assurance of the necessity of planning for spare parts and maintenance during the lifetime of the
robot.
Due to mass production and integrated circuit technology, processors
and many types of sensors have declined in both price and size over the
past few years while motors and batteries have enjoyed no such benefits and
remain the most costly and bulky components of a robot system. In order
to minimize the size and cost of a robot, we propose to use ever smaller and
lighter motors and batteries until we find a limit for building the smallest
robots possible. Recent advances in silicon micromachining technology have
brought about the appearance of micromechanical motors. These motors are
on the order of a few hundreds of microns in diameter and are actually etched
on-chip [2,301. One might question the usefulness of such tiny motors, but if
all we wanted to do was to locomote the chip on which they were fabricated,
then we would have a system in which the motors were of the same scale as
the sensors and processors. Putting an entire robot system on-chip would
allow for mass production using IC fabrication technology and costs would
plummet.
2
Applications of Gnat Robots
Our concept of mobile robots centers around applications that we can imagine for our traditional model of a mobile robot - usually some large, klunky,
heavy, slow, dumb machine. However, if we imagine one-chip robots with
sensors, computers, actuators and power supplies all on one chip, our concept of robotics becomes completely different.
At present, we probably never think of building any machine at a size
smaller than our thumb. This is because we can't build a machine smaller
than the motor we have in it and the smallest motors we've ever seen might
be about as big as our thumb knuckle. The problem is that our thinking
has been constrained by our implementation options.
There are many useful tasks which can be carried out by robots in entirely different implementation ways than what we think of now. Furthermore, many tasks which we would never consider automating could be automated if we had completely orthogonal implementations of robots to what
we presently have. We roboticists often seem to fall into the trap of thinking
too anthropomorphically, i.e. of having robots on the same scale as ourselves
and of having sensors and actuators similar to our own.
Imagine a one-chip robot the size of a gnat which is totally autonomous.
It flies, crawls or swims like a boat. It has microsensors on-board which
allow it to see, smell or feel. What could we do with such machines?
First of all, gnat robots can be used as small autonomous sensors able
to get into hazardous or hard to reach places. Have a suspicious O-ring on
your favorite Space Shuttle? Send a little crawling robot laden with sensors
back to check it out.
Microrobots can be used to fix things too. Imagine a break in an electrical conduit somewhere underground. The robot crawls along it, measuring
the conductivity between legs on each step. When one leg steps on nothing,
the open circuit is found. Then it extends that leg further until it steps over
the break and reconnects the circuit. Then it can just stay there forever,
having fixed the problem - as an autonomous piece of wire.
Another application is farming. Gnat robots fly over fields with their
moisture sensors or infrared sensors [25] and determine exactly which sections of the crop need water. Then they swoop down onto a smart valve
which controls the irrigation system and tap a message into the touch sensor on the top of the valve. Such precision control of limited water reserves
could turn deserts into bread baskets.
Space applications are another appealing realm for cheap lightweight
robots. Gnat robots could be used to inspect the outside of the Space
Station, be employed as forward scouts for a Mars rover, or be used as
atmospheric probes for planetary exploration. Their small size and light
weight would enable hundreds to be taken along on any shuttle flight for
very little cost.
Gnat robots would also make excellent sentries. Mobile sensor platforms
could fly around a warehouse checking for any strange disturbances. Their
low cost would allow many to be used, increasing coverage and effectiveness.
We can imagine reconnaissance operations. Nat the Gnat goes flying
down the hallway (he has a map of the floor stored away), turns left into
your friend's office, grabs a snapshot into his CCD array and flies back to
your office where you do a core dump to see what your friend was up to.
We can also imagine little messenger robots, cheap enough that every
outdoorsman carries a handful of them in his pockets. Next time you're
stuck on Mt. Everest in an avalanche you send a few a these on their way
to get you some help. They deliver little electronic messages which you
programmed in from your pocket calculator.
A huge variety of applications for gnat robots can be found if we don't
let our creativity become inhibited by preconceived notions of what a mobile
robot should look like or how it should act. There are plenty of applications
more interesting than sweeping the floor.
Gnat robots fall into two basic categories of usage. First, they can be
used as autonomous sensors, in which case we want to see what it sees,
but we can't get into the places it can. It brings back or transmits back
the information we desire. The explorer robots and sentry robots fall into
this category. The other way that gnat robots can be used is actually as
autonomous robots. Here, we don't care to see what it sees, but rather
that it just carries out the correct action based upon what it sees. It fixes
whatever's broken and man is out of the loop. The advantage is that we
don't have to retrieve the robot to retrieve data. The electrical conduit fixer
and the irrigation robot fall into this category.
3
Technology To Exploit
Is building a gnat robot feasible? Bringing gnat robots out of the realm
of science fiction and into reality can be accomplished if we exploit recent
breakthroughs in technology in five major areas: microsensors, control systems, actuators, power sources and systems integration.
Microsensors are needed so that the gnat robot can perceive its surroundings and react accordingly. In order to build an integrated robot, we need
to be able to manufacture sensors in a manner compatible with traditional
IC processes. This ensures low cost of mass production and the ability to
put computational circuitry on the same chip as the sensor. Integrating sensors and circuitry provides greater sensitivity since it removes the possibility
of introducing noise or parasitic capacitances which would otherwise occur
in chip to chip wiring. The elimination of this interconnect also facilitates
smaller packaging. Many types of microsensors are already available. Solid
state imaging devices such as black and white CCD cameras have been on
the market for some time with color cameras beginning to emerge. Research
in computer vision has led to algorithms appropriate for mobile robot navigation using these types of sensors, such as one-line stereo [26]. The military
has been using thermal imaging devices for years and they have active research programs ongoing in infrared focal plane arrays (20]. Recently, silicon
micromachining techniques, in which actual physical structures are etched
on chip, have produced a host of new sensors [23]. Chemical sensors [18],
pressure sensors [10] and thermopile infrared sensors [9] are just a few of the
types of sensors being developed in this rapidly expanding field.
How a robot manages to survive and carry out useful tasks is the responsibility of the control system. Taking sensory input and converting it
to appropriate actuator commands is an active area of research, and the
MIT mobile robot group has been developing a control system with some
new and innovative approaches [6]. The idea is to build a system based on
lots of simple finite state machine modules sending messages to each other.
This distributed control system, called the subsumption architecture, is built
in layers with each succeeding layer incorporating more complex behavior.
Lower layers which are built first and debugged, continue to work should
higher levels break, granting robustness to the system. These lower level
behaviors mimic insect level intelligence, incorporating behaviors such as
running away, following, etc. Studies of insect control systems have shown
a remarkably similar architecture [8]. A central tenet of the subsumption
architecture is that there is no central control; no one module acts as a master keeping a representation of the environment and doing planning from
a global map. Thus no single processor can create a bottleneck. The idea
is to achieve intelligence without representation [5]. Indeed, much of the
intelligence resides in the peripherals. This too, is similar to insect control
systems [19]. The beauty of this approach is that each module is very simple and in fact can be compiled down to chip level. A silicon compiler is
presently being developed to do just that [29].
Whatever form of locomotion the gnat robot utilizes, whether it be flying, crawling, swimming or rolling, it will need actuators. Fabricating motors on chip is an area where technological breakthroughs are right around
the corner. Although silicon's electrical properties have been well known and
developed, the mechanical properties of silicon are just now beginning to be
exploited. Surprisingly, silicon has a tensile yield strength three times higher
than that of stainless steel wire (although it yields by fracturing instead of
deforming plastically like most metals) [23]. High resolution planar lithography for defining circuit layouts can be used to create precisely defined
miniature mechanisms. Chemical etching can then sculpt three-dimensional
features. A wide variety of devices utilizing mechanisms such as cantilever
beams and thin membranes have been developed to create vapor and pressure sensors [1,22,16]. Now however, people are going one step further and
building miniature mechanisms such as gears and motors to do mechanical
work [30,2]. It turns out that at the very small feature sizes achievable with
micromachining, electrostatic fields can develop gap energy densities of the
same order of magnitude as the gap energy densities in macroscopic, conventional magnetic motors. Such an electrostatic motor has been conceived that
incorporates a freely rotating disk etched in place between two stators [2].
A synchronous variable capacitance motor with a 300rpm radius, its no-load
power output has been calculated to be 1.8mW with a maximum angular
velocity of 600,000rpm. A power transmission system for this motor has
yet to be designed, however.
Coupling enough power to the gnat robot to drive the motors and circuitry and yet requiring small size and low weight poses some significant
challenges. Total capacitance available on chip is extremely small and very
little power can be stored on-board. Battery technology has never scaled
in the way microelectronics has and piggy backing an off-the-shelf battery
to a gnat robot would cost far more in terms of weight than the rest of the
system. Laying down batteries on chip might be one possibility, eliminating
the weight of the package. Another possible alternative is to radiate power
and not store any on-board, but use it up continuously. Some work in this
area has been done with magnetic power coupling but requires the magnets
to be within a few millimeters of the chip [26]. Solar cells seem promising
because they can be made from single crystal silicon and therefore would be
low weight and compatible with an IC process. Solar cells have problems
with leakage currents when placed in series however, and in addition are typically inefficient, running in the 10% range. Recently, double-sided cells have
yielded improved performance [17] and single-sided cells have been reported
with efficiencies of 22% [27]. Further improvements in solar cell technology
have yielded 24% efficiency[28].
Finally, system integration issues have to be addressed and myriads of
tradeoffs made. Putting all the subsystems together and coming up with
appropriate sensor technology to match a useful means of locomotion is an
intriguing problem. Tradeoffs will have to be made between available real
estate on-board chip, power usage and degrees of complexity of behavior
attainable. Of course, building a large scale mobile robot involves many of
these same constraints and we've gained -a lot of experience in this area of
putting systems together [7].
4
Design Strategies for Gnat Robots
Scaling down all the way to chip level, however, adds a new world of constraints. How will the constraints of small size and low weight affect our design choices? There are two major guiding considerations. The first is that
we want to minimize power consumption. Very little power consumption
means not only that we can use a smaller power source, but that alternative
power sources might become feasible. For instance, some digital watches are
made to run on such low power these days that they are actually powered
by water batteries [14]. The second guiding consideration is that we want
to save chip real estate. If we maximize on-chip integration we will need less
interconnect and the resulting package will become smaller and lighter.
The problems associated with integrating all the necessary subsystems
into a gnat robot lead to a set of design strategies that will most effectively
deal with the constraints of minimizing power consumption and maximizing
level of integration.
The first strategy is to use all passive sensing in order to conserve power.
Right now on the MIT mobots we use sonar, infrared and light-striping for
ranging and proximity sensing. Laser rangefinders are another commonly
used sensor for mobile robots. All of these sensors radiate energy and detect
a return signal. However with our low-power gnat robot, we don't have the
luxury of throwing away power in this way. What we can use instead are
passive imaging devices such as cameras and thermal sensors. Additionally,
we can experiment with other passive sensors such as electrostatic or capacitive proximity sensors, compasses, autofocus mechanisms and chemical
sensors.
The second strategy is to be clever about using sensors so that one sensor
can be used to sense more than one physical phenomenon. This is motivated
by the necessity to conserve chip real estate. Presently the MIT mobots
use a different sensor to sense each different type of physical phenomenon
needed. Sometimes multiple sensors are used to improve the accuracy of a
single type of reading [13]. These strategies are fine for a macroscopic robot
but unworkable for an integrated gnat. Researchers at the MIT Biomedical
Engineering Center have developed a sensor array that uses one sensor to
measure more than one thing and uses two sensors in combination to measure
parameters that neither sensor alone could measure [11]. They have a tumor
probe with an array of ten oxygen-temperature sensor pairs along it. The
temperature sensors not only give an indication of the thermal gradient but
also blood flow and the thermal properties of diffusivity and conductivity.
The oxygen sensor, in conjunction with the temperature sensor, provides
a measure of the partial pressure of oxygen. This very interesting scheme
can be carried over to a sensor system appropriate for a gnat robot. On
our gnat, we can combine an infrared sensor array with ranging techniques
normally used with CCD cameras, such as stereo and motion algorithms.
In this way we get both heat signatures and range data from one piece of
hardware. This enables us to perform tasks such as obstacle avoidance and
warm-body tracking with the same amount of physical hardware that we
would have previously used to do just stereo.
The third strategy is to carry the idea of doubling up sensors one step
further and actually use actuators as sensors. The micromotors will be
sensitive to changes in air pressure and so can be used as pressure sensors.
Furthermore, the motor control system that keeps the micromotor's rotor
spinning level deals with counteracting the torques generated on the rotor
as the gnat changes its orientation. Thus these motors can be used as gyroscopes and provide heading and orientation information. These gyroscopes
hold the possibility of becoming as accurate as optical or mechanical gyros
while costing orders of magnitude less.
Finally, in order to minimize power consumption, we'll have to use the
most efficient power transmission system possible. Gearing down these motors to provide useful torque output will be one of the most formidable
problems and large losses here could wipe out gains made by other strategies. Very little is known about power transmission systems at such small
scales and what the effects of friction, viscosity, small size and high speeds
will have on transmission efficiencies.
5
Details of a Proposed Implementation
Where should we start in our quest to build a gnat robot? Certainly there
are many technical issues that have to be tackled. What we need is a
design for a first prototype, based on sound engineering choices, which can
demonstrate the feasibility of the general concept. Can we show that the
numbers we obtain for motor and power supply output are in the ballpark
of the locomotion requirements of the gnat robot? What would be the best
choices for the first chassis, sensor system, power supply and control system?
Let's start with locomotion options. For a swimming, crawling or rolling
robot just about everything is a big obstacle compared to a gnat. Furthermore, crawling robots would require all the problems of legged locomotion to
be solved first. The resulting complex control system with the ensuing large
number of motors necessary to run all the joints is a major research project
in itself. A swimming locomotion system for a gnat robot also provides
problems. As a body is scaled down in size, viscous forces have relatively
more effect than inertial forces since viscous forces are proportional to the
surface area of a body and fall as the square, whereas inertial forces are proportional to the mass of a body and fall as the cube. This is why amoebas
are often said to effectively swim through asphalt. Flying too, has its problems with viscosity. Small airplanes have low Reynold's numbers - the ratio
of inertial to viscous forces. Tiny insects do manage to fly however, and
so by existence proof we should be able to make a gnat robot fly. Flying
has advantages because it overcomes many of the obstacle avoidance and
complex control systems issues but very little is known about flight at low
Figure 1: A rubber powered airplane with a 50mg airframe and 30mg rubber
band - Mark Drela.
Reynold's numbers, especially flapping flight.
Rather than try and solve the mysteries of insect flight right now, I
propose we find the smallest and lightest known aircraft available today and
see if we can make it autonomous; powered by onboard motors and power
supply and controlled by onboard sensors and control systems. This way,
we can work on the scaling issues involved in the propulsion system and
intelligence areas without having to break new frontiers in aerodynamics.
Probably the record for such an aircraft is held by Professor Mark Drela
of the MIT Aeronautics and Astronautics Department. His rubber band
powered craft which is approximately five inches long, is shown in Figure 1.
It weighs 80mg including the 30mg rubber band.
This remarkable airplane, which can stay aloft for six minutes, has an
airframe made out of splinters of balsa wood and a wing composed of a
film I1m thick [12]. The plane is inherently stable and a slight twist of the
wing causes it to fly in circles. The upswept wingtips keep the plane level
while turning, alleviating the need for ailerons. This feature means that if
we were to automate this airplane, only one additional actuator would be
needed for control - an actuator to steer a rudder. Climbing is accomplished
by increasing the speed of the propeller. Some of the characteristics of the
Table 1: Flight parameters of the rubber powered airplane - Mark Drela.
Weight of the airframe
Weight of the rubber band
Velocity
Drag
Power delivered by prop = dragxvelocity
Overall efficiency
Power delivered by rubber band
Angular velocity of the prop
Torque applied to the prop
Reynold's number
Lift:drag
Energy density of the rubber band
50mg
30mg
1ý = .31 = .7mph
0.4g force = .00039N
.12mW
17 = .3
.4mW
200rpm = 21 rada
191sNm
800- 1000
2.5
15W
59_
airplane are shown in Table 1.
What's intriguing here, is that the power supplied by the rubber band,
.4mW, is less than the no-load power output, 1.8mW, of the MIT micromotor mentioned earlier. Clearly it seems plausible then, that micromotors
of this type might be adequate to power this aircraft. With the addition of
high efficiency solar cells to supply power, and passive sensors and a control
system for obstacle avoidance, we just might be able to build a very small
airplane that could fly autonomously.
The MIT micromotor, shown in Figure 2 (from [2]), has another characteristic that makes the feasibility of powering the plane with this motor
look more promising. Because of its extremely small size, its mass is very
low. A silicon micromotor approximately 7001im in diameter on a 500pum
thick wafer would weigh .4mg, roughly two orders of magnitude less than
the weight of the rubber band.
The airplane's propeller requires an applied torque of 19pNm at an angular velocity of 200rpm, equivalent to .4mW of power. The corresponding
no-load figures for the micromotor [2] are shown in Table 2.
Since at the present time, the design has no provision for attaching a
load, the maximum angular velocity, wma, is determined by setting the
electric drive torque, re, equal to the viscous torque, 7,, where all the power
-p______
Figure 2: The synchronous variable capacitance rotary micromotor [2].
Table 2: Performance of the MIT micromotor under no load (2].
re
r,
dmaz
Pa =
"W.ma,
Drive torque
Viscous torque at max velocity
Max angular velocity
Power delivered
2.9 x 10-8Nm
2.9 x 10- 8 Nm
6.2 x 104 rd = 600,000rpm
1.8mW
-82.
x
2.9 x 10 Nm
W
Wmax
4 rads
6.2 x 10 a
Figure 3: Electric drive torque and viscous torque.
is dissipated by the rotor churning up the air. The drive electronics would
then have to also switch at this speed in order for the motor to work.
What we need to know however, is what speed we'll have to operate at in
order to get useful torque output. It turns out that this type of motor has a
constant 7, for all values of w [3,4,31]. Since r, increases linearly with w, the
useful torque, rL, is the difference between these two curves and decreases
linearly with omega as shown in Figures 3 and 4.
Since the power delivered to the load, PL, is equal to the product rLw,
maximum efficiency is derived from the motor when it is operated at Wop =
Wiaa and r•, = . These values are shown in Table 3. Note that the power
output is still roughly equal to that of the rubber band. The torque-velocity
relationship however, is clearly in the wrong ratio. We need to increase the
motor's torque 1,270 times and decrease its speed by a factor of 1500 in
order to drive the propeller. A very efficient gearing system is needed since
we don't have much power to spare.
Assuming some transmission system can be developed, we turn now to
the question of providing power. Probably the highest power density solar
cell developed so far is a double sided design which utilizes photons captured
by reflection from a surface below the cell to boost the specific power to
700ý [17,15]. Thus the .5mW of power necessary to drive the motor at its
operating point can be achieved with .7mg of silicon. For an 80pm thick
cell and p,i = 2300 4', this takes up only 3.5mm 2 of space. The micromotor
-8
2.9 x 10 Nm
-8
1.5 x 10 Nm
W
_ax
Wmax
4
2 4 rad
rads
3.1x 10
6.2 x 10 300,000rpm 600,000rpm
Figure 4: Useful torque that can be delivered to a load.
Table 3: Operating Point for the MIT micromotor.
PL
qPL
Torque available for the load
Operating velocity
Power delivered to the load
1.5 x 10-'Nm
3.1 x 104
.46mW
= 300,000rpm
needs 100V applied across the capacitors formed between the stators and
the rotor. At 0.6V per solar cell, 167 cells need to be run in series to provide
100V. Each cell would then be 0.02mm
m.2 The total weight of the propulsion
system, motor plus solar cell, is only 1.1mg. This is still more than an order
of magnitude less than the weight of the rubber band.
It seems plausible then, that continuous flight can be achieved and we
need not be constrained by the finite energy of a rubber band. However, this
example of powering the airplane by a micromotor of the type being worked
on at MIT is not intended to give the impression that these motors are
widely available or even that this particular design is the right choice. The
MIT micromotor work described here is, at this time, still at the design stage
and no motors have yet been fabricated or tested. The numbers used in this
paper for maximum torque and maximum angular velocity are theoretical
limits and are not meant to imply actual working values. Their inclusion in
this paper is merely to show that calculated power outputs of these types of
motors are in the same ballpark as those needed to propel a small vehicle.
What sort of sensors and control system can we put onboard to automate this plane? We are presently developing a sensor system based on a
pair of single scanline CCD cameras to implement obstacle avoidance algorithms using stereo and motion techniques. Some novel ideas regarding
automatic calibration of the cameras allow the algorithms to work without
initialization, which usually requires the use of a test pattern. Thus robustness is provided should the cameras become misaligned while in use.
Stereo provides information regarding bearing and range to a target. This
is accomplished by matching vertical edges in two images and then triangulating. This assumes however, that the optical axes of the two cameras are
perpendicular. Errors here can be calibrated out if a few initial distances
are known, and that information can be provided by motion algorithms
which match feature points from one camera that has moved in position
over time. This type of passive sensor system can be implemented on chip
and incorporated into a very small package. Several such systems can be
used to provide both forward and peripheral vision and enable the airplane
to discern placement of obstacles.
Converting stereo and motion information to appropriate rudder and
propeller commands can be accomplished through a subsumption architecture control system very similar to the ones presently running the MIT
mobile robots [6]. Figure 5 shows a level zero version of a control system
for a miniature airplane. The camera module provides a local map in robotcentered polar coordinates of bearing and range to obstacles and sends this
airplane
rudder
Figure 5: Runaway behavior. Level 0 control system for an airplane.
map to the feelforce module. The feelforce module does a vector sum on this
map, treating distant obstacles as weakly repulsive forces and close obstacles
as strongly repulsive forces and produces one resultant vector that signifies
the path of least resistance. The runaway module monitors this force and
when it becomes significant it sends a command to the rudder, directing the
airplane to turn left or right. The overall behavior achieved here is that the
airplane will fly in a straight line and turn before it collides with an obstacle.
A slightly more sophisticated behavior can be achieved by developing a
level 1 control system. Level 0 is left unchanged, but augmented by new
modules that sometimes supersede level 0 modules. Different levels are
combined by means of a suppressor mechanism in which outputs from higher
level modules can suppress outputs from modules in lower levels. A time
constant is associated with each suppressor node so that suppression occurs
for a fixed period of time after the higher level module fires. Figure 6
shows the level 1 control system which implements a behavior in which
the airplane will randomly change direction, and in addition, will avoid
obstacles. The wander module provides new headings at discrete increments
in time and sends them to the avoid module. The avoid module receives
both the resultant force vector and the heading command, and if necessary,
computes a slightly perturbed heading so as to avoid any obstacles. It
supplies this result to the rudder and propeller, directing the airplane to
turn left or right, climb or dive. The output of the runaway module is
suppressed for the amount of time set at the suppressor node. Each module
in this control system is a simple finite state machine and can be compiled
into silicon directly.
The example of automating the miniature airplane here, is not an abandonment of the goal to build a gnat-sized robot, but rather is intended to
rudder
airplane
command
propeller
Figure 6: Wander behavior. Level 1 control system.
give concrete evidence of feasibility of integrating a miniature propulsion
system with a sensor suite and a subsumption architecture control system,
while temporarily alleviating the need for addressing the problem of aerodynamic flight at low Reynold's numbers. The driving issue is cost. If all
these things could be integrated and mass produced, then the price of robot
systems would drop drastically and open up new areas of application.
What this example has tried to show is that it is worthwhile to take a
different course than that which is typically taken in mobile robotics today.
That is, it is worthwhile to pursue the idea of using ever smaller and lighter
motors and power supplies with the aim of developing cheaper robots, which
then hopefully will make robotics more useful to society.
6
What We Need From Micromachining Technology
We've highlighted a specific set of technologies that fit together appropriately to implement a microrobotic system. Obviously, some elements are
farther along in development than others. What are the biggest problems
and where do we need to concentrate research so that we can actually build
such systems in the near future?
Micromotor development is probably the primary concern. Methods for
transmitting power need to be studied. A single-stage gear to achieve the
torque reduction required for the airplane would be two meters in diameter.
Clearly there are better solutions. Harmonic drives, multi-stage gears or
perhaps even fluid gears might be appropriate. Very little is known about
fluid dynamics in the scenarios presented by micromotors: boundary layer
domains, high ratios of viscous to inertial forces and incredibly high rotor
speeds. Alternative designs for micromotors could possibly eliminate the
need for gearing altogether. Miniaturization of surface acoustic wave motors,
which have a high torque to mass ratio, may be the answer.
Interfacing micromotors to macroscopic loads is another problem area.
If attaching a propeller to a micromotor requires hand assembly with a very
tiny drop of epoxy, then all the advantages of low cost, mass production
are lost. Etching propeller mounts or building small robots to do micro
assembly may be necessary.
Another technology area that requires research is solar power. Isolation
between cells is very important when putting cells in series on the same chip.
Improved process technology to yield lower leakage currents is needed. An
alternative solution might be to use chemical etching to increase isolation.
Finally, we need research in micro-aerodynamics, a field relatively untouched. It's possible that although we don't know how to make a gnat-sized
robot fly today, development of micromotor technology may be a useful tool
in studying insect flight.
7
How Gnat Robots Will Change Robotics
The low cost of gnat robots will change the way we think about using
robotics in two major ways. First of all, we can begin to think in terms of
massive parallelism. Instead of large machines which have parts from many
vendors and need many interfaces, we use millions of very simple robots to
carry out massively parallel work. Take for example, the job of cleaning
barnacles off the side of a ship. We could use one large, crane-like robot
to lean over the side of the ship and scrub or we could imagine throwing
millions of little gnat robots at the side of the ship and have them munch
away in parallel. The idea is analogous to the use of a Cray vs. a Connection Machine in terms of sequential vs. parallel processing. The crane robot
would also be large and costly and require spare parts and trained personnel
to maintain it. The gnat robot on the other hand, would be small, cheap
and thrown away when it breaks.
This brings us to the second way in which low-cost gnat robots will
change the way we use robot technology. We will see the emergence of
cheap disposable robots. Once designed, gnat robots are produced in one
fabrication process and can be very cheaply manufactured. They can be
thrown away or left behind when they finish their task or run out of power.
Presently, a sentry robot guarding a warehouse just might be the most
expensive thing in that warehouse and we might tend to worry more about
somebody stealing our robot than any merchandise. Gnat robots will come
to be regarded like ball point pens. They'll be cheap enough that we just
toss them out and buy new ones when the old ones break.
8
Conclusion
A new direction in robotics is needed in order to reduce costs so that robots
can become more useful to society. Integration of an entire robotic system
onto a single piece of silicon is the avenue we need to pursue. By integrating
propulsion systems, sensors and intelligent control, mass production and low
cost can be achieved, opening up a new range of applications. Instead of
spending energy on large, mainframe robots which will never be cheaply
mass produced, we should focus research on microrobots. Just as declining
costs enabled personal computers to become widely available, so too will
declining costs enable robot technology to become a part of everyday life.
This paper has argued that such machines are not science fiction, but are
in fact, well within the realm of feasibility. We've outlined which technology
areas and problems need to be addressed and have outlined some strategies
for design that will enable this technology to come to fruition in the fastest
possible manner.
The example of automating the miniature airplane was designed to illustrate specifically how the appropriate pieces could be combined to realize
a microrobotic system. The framework of a lightweight airplane was chosen because it provides the simplest scenario for implementing a complete
system, due to the fact that it requires very little power input and can be
automated with a straightforward control system. The point of this example
was not to imply that the micromotors mentioned would be the right choice
for design of a propulsion system, or that an airplane of a macroscopic scale
would be acceptable. Rather the goal was to put all the pieces in the concrete context of a system in order to demonstrate feasibility, to show that
this direction is worthy of pursuit, and to pinpoint which problems need to
be solved.
For instance, although the micromotors discussed in this paper have a
calculated power output roughly equivalent to that necessary to power a
small airplane, we know that higher torque and lower speed are required,
and we can begin to concentrate on those issues. Whether the solution is a
mechanical or hydraulic gearing system, or a totally different kind of motor
altogether, the exercise of putting the problem in the context of an entire
system has been helpful.
All of this is not an abandonment of the ultimate goal of building a truly
gnat-sized robot either, but merely good engineering use of abstraction, i.e.
isolating tough problems. By separating the problem of microaerodynamics
from integration of miniature propulsion, sensor and control systems, it
might be possible to solve them iteratively. Advances in one area may
provide tools for research in the other. Hopefully in this way, we can make
steady progress in solving the many problems involved in building gnat
robots.
9
Acknowledgments
Discussions with numerous people spawned many of the ideas in this paper. The author would especially like to thank Stephen Bart and Jeff Lang
of the MIT Laboratory of Electromagnetic and Electronic Systems; Roger
Howe and Martin Schlecht of the MIT Microsystems Technology Laboratories; Mark Drela, Steve Bussolari, Juan Cruz, Harold Youngren and Steve
Finberg from the MIT Daedalus Project; David Staelin and Greg Wornell,
MIT Research Laboratory of Electronics; Mike Curley, MIT Biomedical Engineering Center; Harry West, MIT Mechanical Engineering Design Center
and Ken Salisbury, Ken Pasch, Al Ward, Mike Caine, Carl Ullrich, Steve
Eppinger, Eric Vaaler and Rod Brooks of the MIT Artificial Intelligence
Laboratory.
References
[1] James B. Angell, Stephen C. Terry, and Phillip W. Barth. Silicon
micromechanical devices. Scientific American, 44-55, April 1983.
[2] Stephen Bart and Jeff Lang, MIT Laboratory of Electromagnetic and
Electronic Systems; Martin Schlecht and Roger Howe, MIT Microsystems Technology Laboratories. Personal communication.
[3] Stephen Bart. MIT Laboratory of Electromagnetic and Electronic Systems. Personal communication.
[4] B. Boll'e. Electrostatic motors. Philips TechnicalReview, 30(6/7):178194, 1969.
[5] Rodney A. Brooks. Intelligence without representation. In Proceedings
of the Workshop on The Foundationsof AI, Endicott House, June 1987.
Forthcoming.
[6] Rodney A. Brooks. A robust layered control system for a mobile robot.
IEEE Journalof Robotics and Automation, RA-2(1), April 1986.
[7] Rodney A. Brooks, Jon Connell, and A.M. Flynn. A mobile robot with
onboard parallel processor and large workspace arm. In Proceedingsof
AAAI, August 1986.
[8] Jeffrey M. Camhi. The escape system of the cockroach.
American, 243(6):169-171, December 1980.
Scientific
[9] Il Hyun Choi and Kensall D. Wise. A silicon-thermopile based infrared
sensing array for use in automated manufacturing. IEEE Trans. on
Electron Devices, ED-33(1):72-79, 1986.
[10] Samuel K. Clark and K.D. Wise. Pressure sensitivity in anisotropically etched thin-diaphram pressure sensors. IEEE Trans. on Electron
Devices, ED- 26(12):1887-1896, 1979.
[11] M.G. Curley, L.M. Peznola, J.C. Weaver, D.J. Edell, S.K. Burns, and
H.F. Bowman. Multipoint monitoring of temperature, oxygen tension,
perfusion and thermal properties. In N.A. Langrana, editor, Advances
in Bioengineering,1985. ASME Winter Annual Meeting, Miami, Nov.
1985.
[12] Mark Drela. MIT Aeronautics and Astronautics Department. Personal
communication.
[13] A.M. Flynn. Redundant Sensors For Mobile Robot Navigation. Technical Report 859, MIT AI Lab, 1985.
[14] Marty Heitner. As time flows by. IEEE Spectrum, March 1987.
[15] Rudolf Hezel. Institut f'ir Werkstoffwissenschaften VI, Universit•it
Erlangen-Niirnberg, D-8520 Erlangen, West Germany. Personal communication.
[16] R.T. Howe. Polycrystalline silicon microstructures. In C.D. Fung, P.W.
Cheung, W.H. Ko, and D.G. Fleming, editors, Micromachining and
Micropackagingof Transducers,pages 169-187, Elsevier, 1985.
[17] K. Jaeger and R. Hezel. Bifacial metal insulator semiconductor inversion layer solar cells based on low temperature silicon surface passivation. In Proceedings of the 19th IEEE Photovoltaic Specialists Conference, New Orleans, May 1987.
[18] R.G. Kelly and A.E. Owen. Microelectronic ionsensors: a critical survey. IEE Proc., Vol. 132 Pt. 1(5):227-236, 1985.
[19] T. Kleist. Bee navigation: the eyes have it. Science News, 130(14):214,
October 4, 1986.
[20] M.R. Kruer, D.A. Scribner, and J.M. Killiany. Infrared focal plane array technology development for navy applications. OpticalEngineering,
26(3):182-190, March 1987.
[21] Kurt E. Petersen. Dynamic micromechanics on silicon: techniques and
devices. IEEE Transactionson Electron Devices, ED-25, 1978.
[22] Kurt E. Petersen.. Silicon as a mechanical material. In Proc. IEEE,
pages 420-457, 1982.
[23] Kurt E. Petersen. Silicon sensor technologies. In Technical Digest,
pages 2-7, IEEE International Electron Devices Meeting, Washington,
D.C., December 1985.
[24] J. Raloff. Infrared can spy plant stress before eyes do. Science News,
127(5):70, February 2, 1985.
[25] Charles W. Selvidge, Adam C. Malamy, and Lance A. Glasser. Power
and communication techniques for physically isolated integrated circuits. In Proceedingsof the Stanford Conference on Advanced Research
in VLSI, March 1987.
[26] Bruno Serey and Larry Matthies. Obstacle avoidance using 1-d stereo
vision. CMU Technical Report. In preparation.
[27] R.A. Sinton, Y. Kwark, J.Y. Gan, and R.M. Swanson. 27.5-percent silicon concentrator solar cells. Electron Device Letters, EDL-7(10):567569, October 1986.
[28] Richard Swanson. Stanford University. Personal communication.
[29] Markus Toth. A Design Compilerfor the Intelligent Control System of
a Mobile Robot. Master's thesis, MIT AI Lab, August 1987. Forthcoming.
[30] W.S.N. Trimmer and K.J. Gabriel. Design considerations for a practical
electrostatic micro motor. Sensors and Actuators, 11(2):189-206, 1987.
[31] J.G. Trump. Vacuum Electrostatic Engineering. PhD thesis, Massachusetts Institute of Technology, 1933.